Text Lives in the Guidance: Preventing Silent Text Rewriting in Cached Diffusion Transformers
Abstract
Diffusion transformers (DiTs) now render long, legible text in images, and feature caching is a standard way to accelerate them. Caches are judged by image-level metrics, which do not check whether the rendered text is still correct. We find that caching silently rewrites this text. On Z-Image, TaylorSeer at 2.9x fewer FLOPs lowers word accuracy from 62% to 46%, while ImageReward and CLIP Score barely change. Sampling with fewer steps at the same compute keeps almost all of this text, so the loss comes from how the cache predicts the skipped steps. Decomposing the cache error along classifier-free guidance (CFG), we trace this loss mainly to the guidance term, the difference between the conditional and unconditional predictions, which caches extrapolate worst on text. Correcting the guidance inside the text alone recovers most of the lost text. We therefore propose GuideCache, which recomputes both CFG branches only on the tokens with the largest predicted guidance error, using the guidance terms that the cache already stores. At 2.4x fewer FLOPs, GuideCache restores full-compute prompt fidelity on Z-Image and Qwen-Image, and on Z-Image it renders text as well as or better than ToCa, TeaCache, and FasterCache while staying much closer to the full model's output. GuideCache works with caches of any order, and on top of fewer sampling steps it keeps more text than step reduction at a higher speedup. These results suggest that acceleration methods for text-rendering models should be evaluated on the text they render.
Figures & tables
| Qwen-Image | Z-Image | |||||||||||
| Method | Word | Prompt | PSNR | LPIPS | FLOPs | Lat. | Word | Prompt | PSNR | LPIPS | FLOPs | Lat. |
| Full (50 steps) | 0.605 | 0.615 | 0 | 1.00 | 1.00 | 0.621 | 0.665 | 0 | 1.00 | 1.00 | ||
| 17 anchor evaluations (2.94 FLOPs) | ||||||||||||
| Step reduction, 17 steps | 0.519 | 0.579 | 13.5 | 0.417 | 2.94 | 2.87 | 0.587 | 0.666 | 12.9 | 0.421 | 2.94 | 2.91 |
| Step reduction, 22 steps | 0.547 | 0.604 | 15.2 | 0.321 | 2.27 | 2.24 | 0.607 | 0.666 | 14.0 | 0.353 | 2.27 | 2.26 |
| FasterCache ∗ | – | – | – | – | – | – | 0.565 | 0.652 | 15.4 | 0.327 | 1.64 | 1.70 |
| + GuideCache | |||||
| Cache | Anchors | FLOPs | Plain | ||
| Full compute | 50 | 1.0 | 0.645 | ||
| FORA ( ) | 28 | 1.8 | 0.641 | 0.644 | 0.643 |
| 22 | 2.3 | 0.630 | 0.645 | 0.654 | |
| 17 | 2.9 | 0.622 | 0.638 | 0.654 | |
| 13 | 3.8 | 0.475 | 0.634 | 0.645 | |
| Choice | Setting | Word | Prompt | FLOPs |
|---|---|---|---|---|
| Plain cache | 0.386 | 0.601 | 2.94 | |
| Full compute | 0.540 | 0.645 | 1.00 | |
| Selector | Random | 0.420 | 0.623 | 1.79 |
| ToCa attention score | 0.456 | 0.622 | 1.79 | |
| Guidance share | 0.516 | 0.653 | 1.80 | |
| Shrunk score | 0.532 | 0.647 | 1.80 |
Appendix figures & tables13 assets
Supplementary material from the paper’s appendix.
Appendix
| Anchor | Cached | Select | Mem. | Stored | |
| (s) | (s) | (ms) | (GiB) | (GiB) | |
| Z-Image | |||||
| Full compute | 1.27 | – | – | 12.6 | – |
| TaylorSeer | 1.26 | – | – | 12.6 | – |
| + GuideCache | 1.27 | 0.51 | 0.53 | 16.4 | 0.007 |
| Qwen-Image | |||||
| Quantity | Measurement |
|---|---|
| Setting | Z-Image, 34 of 50 evaluations on the full-compute trajectory (23–24 dB); anchor interval 6 for orders 0–1, 3 for order 2; 48 prompts for pixels, 12 English and 12 Chinese inside the network |
| Pixel error, text / control | 0.96–1.13 (text / background: 3–4 ) |
| Latent noise of equal norm | box fidelity of text within 0.02 of the control up to 17% of the latent norm; 0.853 vs. 0.920 at 33%; control SSIM 0.98 to 0.78 |
| Guidance share | text 0.50 vs. control 0.32 (1.66 ); English 1.55 , Chinese 1.77 |
| Per-token rescaling (Qwen-Image) | TaylorSeer at 17 anchors loses 0.084 word accuracy and 0.018 prompt fidelity with the rescaling, and 0.169 and 0.050 without it (160 prompts, two seeds) |
| Extrapolation error of , text / control | reuse 1.83 , first order 1.96 , second order 1.30–1.38 |
| Exact part | Region | Prompt fid. | Box fid. | Word acc. |
|---|---|---|---|---|
| Guidance term | text boxes | 83% | 42% | 90% |
| Shared part | text boxes | 11% | 2% | 2% |
| Guidance term | control | |||
| Both | text boxes | 108% | 45% | 84% |
| Qwen-Image | Z-Image | |||||||||||
| Method | Word | Prompt | PSNR | LPIPS | FLOPs | Lat. | Word | Prompt | PSNR | LPIPS | FLOPs | Lat. |
| Stacking on a 25-step schedule (3.1–4.2 FLOPs) | ||||||||||||
| Step reduction, 15 steps | 0.485 | 0.569 | 13.0 | 0.456 | 3.33 | 3.20 | 0.565 | 0.663 | 12.5 | 0.452 | 3.33 | 3.30 |
| 25 steps + TaylorSeer (12 anchors) | 0.512 | 0.597 | 15.9 | 0.295 | 4.17 | 3.90 | 0.461 | 0.620 | 13.6 | 0.380 | 4.17 | 4.10 |
| 25 steps + TaylorSeer (16 anchors) | 0.526 | 0.604 | 16.2 | 0.280 | 3.12 | 2.98 | 0.551 | 0.657 | 14.4 | 0.334 | 3.12 | 3.10 |
| 25 steps + TaylorSeer (12) + GuideCache ( ) | 0.554 | 0.611 | 16.1 | 0.288 | 3.71 | 3.36 | 0.599 | 0.661 | 13.9 | 0.358 | 3.63 | 3.50 |
| Qwen-Image | Z-Image | |||||||||||
| Method | Word | Prompt | PSNR | LPIPS | FLOPs | Lat. | Word | Prompt | PSNR | LPIPS | FLOPs | Lat. |
| TaylorSeer | 0.521 | 0.597 | 21.2 | 0.130 | 2.94 | 2.76 | 0.464 | 0.625 | 15.0 | 0.306 | 2.94 | 2.90 |
| + GuideCache 5% | 0.571 | 0.615 | 22.3 | 0.115 | 2.68 | 2.22 | 0.586 | 0.664 | 15.7 | 0.275 | 2.64 | 2.50 |
| + GuideCache 10% | 0.583 | 0.618 | 23.1 | 0.104 | 2.46 | 2.11 | 0.606 | 0.670 | 16.2 | 0.256 | 2.41 | 2.28 |
| + GuideCache 20% | 0.588 | 0.617 | 24.3 | 0.083 | 2.14 | 1.91 | 0.608 | 0.670 | 16.9 | 0.229 | 2.07 | 1.97 |
| + GuideCache 30% | 0.585 | 0.616 | 25.3 | 0.070 | 1.89 | 1.71 | 0.612 | 0.668 | 17.8 | 0.198 | 1.81 | 1.70 |
| Qwen-Image | Z-Image | |||
|---|---|---|---|---|
| Method | Rewr. | New | Rewr. | New |
| Step reduction, 17 steps | 22.2 | 14.3 | 14.2 | 16.8 |
| Step reduction, 22 steps | 15.6 | 11.8 | 10.4 | 14.4 |
| FasterCache ∗ | – | – | 13.3 | 8.2 |
| ToCa ∗ | – | – | 12.1 | 17.2 |
| TeaCache ∗ | – | – | 12.4 | 12.8 |
| Method | Word | Prompt | PSNR | LPIPS | FLOPs | Rewr. |
|---|---|---|---|---|---|---|
| TaylorSeer | 0.464 | 0.625 | 15.0 | 0.306 | 2.94 | 27.7 |
| + GuideCache ( ) | 0.606 | 0.670 | 16.2 | 0.256 | 2.41 | 8.2 |
| HiCache | 0.491 | 0.639 | 15.2 | 0.298 | 2.94 | 22.7 |
| + GuideCache ( ) | 0.596 | 0.672 | 16.4 | 0.255 | 2.49 | 9.7 |
| Full (50 steps) | 0.621 | 0.665 | 0 | 1.00 | – |
| Setting | Word | Prompt | LPIPS | FLOPs | |
|---|---|---|---|---|---|
| 0 ( only) | 0.508 | 0.629 | 0.259 | 2.30 | |
| 0.25 | 0.508 | 0.621 | 0.252 | 2.33 | |
| 0.5 | 0.521 | 0.632 | 0.251 | 2.41 | |
| 0.75 | 0.465 | 0.631 | 0.251 | 2.50 | |
| 1 ( only) | 0.471 | 0.638 | 0.259 | 2.58 | |
| 6, TaylorSeer | 0.442 | 0.616 | 0.286 | 2.94 |
| + GuideCache | ||||||
| Cache | Anchors | FLOPs | PSNR | Plain | ||
| Full compute | 50 | 1.0 | 0.636 / 0.653 / 0.634 | |||
| FORA ( ) | 28 | 1.8 | 23.9–24.3 | 0.635 / 0.648 / 0.640 | 0.638 / 0.650 / 0.631 | 0.637 / 0.649 / 0.632 |
| 22 | 2.3 | 19.7–19.9 | 0.617 / 0.643 / 0.632 | 0.629 / 0.661 / 0.642 | 0.646 / 0.661 / 0.640 | |
| 17 | 2.9 | 16.0–16.3 | 0.632 / 0.613 / 0.608 | 0.631 / 0.645 / 0.670 | 0.648 / 0.660 / 0.659 | |
| 13 | 3.8 | 14.8–14.9 | 0.472 / 0.478 / 0.445 | 0.612 / 0.655 / 0.627 | 0.624 / 0.666 / 0.633 | |
| Z-Image | Qwen-Image | ||||||||||||||
| Box fidelity | Prompt fidelity | Box fidelity | Prompt fidelity | ||||||||||||
| Stratum | Group | TS | +GuideCache | TS | +GuideCache | TS | +GuideCache | TS | +GuideCache | ||||||
| Font size | small | 213 | 0.462 | 0.574 | +0.113 | 0.401 | 0.469 | +0.068 | 106 | 0.591 | 0.646 | +0.055 | 0.486 | 0.517 | +0.031 |
| medium | 213 | 0.757 | 0.813 | +0.056 | 0.708 | 0.736 | +0.028 | 106 | 0.702 | 0.751 | +0.049 | 0.608 | 0.625 | +0.018 | |
| large | 213 | 0.813 | 0.868 | +0.055 | 0.770 | 0.809 | +0.040 | 107 | 0.727 | 0.742 | +0.015 | 0.701 | 0.716 | +0.015 | |
| Script | Latin | 319 | 0.671 | 0.743 | +0.073 | 0.643 | 0.678 | +0.035 | 160 | 0.691 | 0.724 | +0.033 | 0.597 | 0.610 | +0.013 |
| Method | IR | CLIP | PSNR | SSIM | LPIPS | FLOPs |
|---|---|---|---|---|---|---|
| Full (50 steps) | 0.936 | 32.30 | 1 | 0 | 1.00 | |
| TaylorSeer ( ) | 0.763 | 31.90 | 17.6 | 0.715 | 0.315 | 2.94 |
| + GuideCache | 0.888 | 32.38 | 20.7 | 0.783 | 0.219 | 1.98 |
Diffusion Model Caching
Diffusion model caching reuses features, activations, or attention states across denoising steps to reduce inference cost while preserving output quality.
See all 124 papers